X-Git-Url: http://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb2%2Fhash.c;h=18f6b1720eb662b2a0f140a6d23f7c844269b223;hp=0cc93760d1613fa44c4a2681133f27aeadb50c19;hb=a42bba8ec446284256a7c9146ba3525404de474c;hpb=6804501c350181dea8f531142b28c620b70edbd9 diff --git a/ccan/tdb2/hash.c b/ccan/tdb2/hash.c index 0cc93760..18f6b172 100644 --- a/ccan/tdb2/hash.c +++ b/ccan/tdb2/hash.c @@ -42,17 +42,19 @@ uint64_t tdb_hash(struct tdb_context *tdb, const void *ptr, size_t len) uint64_t hash_record(struct tdb_context *tdb, tdb_off_t off) { - struct tdb_used_record pad, *r; + const struct tdb_used_record *r; const void *key; uint64_t klen, hash; - r = tdb_get(tdb, off, &pad, sizeof(pad)); + r = tdb_access_read(tdb, off, sizeof(*r), true); if (!r) /* FIXME */ return 0; klen = rec_key_length(r); - key = tdb_access_read(tdb, off + sizeof(pad), klen, false); + tdb_access_release(tdb, r); + + key = tdb_access_read(tdb, off + sizeof(*r), klen, false); if (!key) return 0; @@ -76,6 +78,30 @@ static uint32_t use_bits(struct hash_info *h, unsigned num) return bits(h->h, 64 - h->hash_used, num); } +static bool key_matches(struct tdb_context *tdb, + const struct tdb_used_record *rec, + tdb_off_t off, + const struct tdb_data *key) +{ + bool ret = false; + const char *rkey; + + if (rec_key_length(rec) != key->dsize) { + add_stat(tdb, compare_wrong_keylen, 1); + return ret; + } + + rkey = tdb_access_read(tdb, off + sizeof(*rec), key->dsize, false); + if (!rkey) + return ret; + if (memcmp(rkey, key->dptr, key->dsize) == 0) + ret = true; + else + add_stat(tdb, compare_wrong_keycmp, 1); + tdb_access_release(tdb, rkey); + return ret; +} + /* Does entry match? */ static bool match(struct tdb_context *tdb, struct hash_info *h, @@ -83,38 +109,33 @@ static bool match(struct tdb_context *tdb, tdb_off_t val, struct tdb_used_record *rec) { - bool ret; - const unsigned char *rkey; tdb_off_t off; - /* FIXME: Handle hash value truncated. */ - if (bits(val, TDB_OFF_HASH_TRUNCATED_BIT, 1)) - abort(); - + add_stat(tdb, compares, 1); /* Desired bucket must match. */ - if (h->home_bucket != (val & TDB_OFF_HASH_GROUP_MASK)) + if (h->home_bucket != (val & TDB_OFF_HASH_GROUP_MASK)) { + add_stat(tdb, compare_wrong_bucket, 1); return false; + } /* Top bits of offset == next bits of hash. */ if (bits(val, TDB_OFF_HASH_EXTRA_BIT, TDB_OFF_UPPER_STEAL_EXTRA) != bits(h->h, 64 - h->hash_used - TDB_OFF_UPPER_STEAL_EXTRA, - TDB_OFF_UPPER_STEAL_EXTRA)) + TDB_OFF_UPPER_STEAL_EXTRA)) { + add_stat(tdb, compare_wrong_offsetbits, 1); return false; + } off = val & TDB_OFF_MASK; if (tdb_read_convert(tdb, off, rec, sizeof(*rec)) == -1) return false; - /* FIXME: check extra bits in header? */ - if (rec_key_length(rec) != key->dsize) + if ((h->h & ((1 << 11)-1)) != rec_hash(rec)) { + add_stat(tdb, compare_wrong_rechash, 1); return false; + } - rkey = tdb_access_read(tdb, off + sizeof(*rec), key->dsize, false); - if (!rkey) - return false; - ret = (memcmp(rkey, key->dptr, key->dsize) == 0); - tdb_access_release(tdb, rkey); - return ret; + return key_matches(tdb, rec, off, key); } static tdb_off_t hbucket_off(tdb_off_t group_start, unsigned bucket) @@ -123,22 +144,88 @@ static tdb_off_t hbucket_off(tdb_off_t group_start, unsigned bucket) + (bucket % (1 << TDB_HASH_GROUP_BITS)) * sizeof(tdb_off_t); } -/* Truncated hashes can't be all 1: that's how we spot a sub-hash */ bool is_subhash(tdb_off_t val) { - return val >> (64-TDB_OFF_UPPER_STEAL) == (1<> TDB_OFF_UPPER_STEAL_SUBHASH_BIT) & 1; +} + +/* FIXME: Guess the depth, don't over-lock! */ +static tdb_off_t hlock_range(tdb_off_t group, tdb_off_t *size) +{ + *size = 1ULL << (64 - (TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS)); + return group << (64 - (TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS)); +} + +static tdb_off_t COLD find_in_chain(struct tdb_context *tdb, + struct tdb_data key, + tdb_off_t chain, + struct hash_info *h, + struct tdb_used_record *rec, + struct traverse_info *tinfo) +{ + tdb_off_t off, next; + + /* In case nothing is free, we set these to zero. */ + h->home_bucket = h->found_bucket = 0; + + for (off = chain; off; off = next) { + unsigned int i; + + h->group_start = off; + if (tdb_read_convert(tdb, off, h->group, sizeof(h->group))) + return TDB_OFF_ERR; + + for (i = 0; i < (1 << TDB_HASH_GROUP_BITS); i++) { + tdb_off_t recoff; + if (!h->group[i]) { + /* Remember this empty bucket. */ + h->home_bucket = h->found_bucket = i; + continue; + } + + /* We can insert extra bits via add_to_hash + * empty bucket logic. */ + recoff = h->group[i] & TDB_OFF_MASK; + if (tdb_read_convert(tdb, recoff, rec, sizeof(*rec))) + return TDB_OFF_ERR; + + if (key_matches(tdb, rec, recoff, &key)) { + h->home_bucket = h->found_bucket = i; + + if (tinfo) { + tinfo->levels[tinfo->num_levels] + .hashtable = off; + tinfo->levels[tinfo->num_levels] + .total_buckets + = 1 << TDB_HASH_GROUP_BITS; + tinfo->levels[tinfo->num_levels].entry + = i; + tinfo->num_levels++; + } + return recoff; + } + } + next = tdb_read_off(tdb, off + + offsetof(struct tdb_chain, next)); + if (next == TDB_OFF_ERR) + return TDB_OFF_ERR; + if (next) + next += sizeof(struct tdb_used_record); + } + return 0; } /* This is the core routine which searches the hashtable for an entry. * On error, no locks are held and TDB_OFF_ERR is returned. - * Otherwise, hinfo is filled in. + * Otherwise, hinfo is filled in (and the optional tinfo). * If not found, the return value is 0. * If found, the return value is the offset, and *rec is the record. */ tdb_off_t find_and_lock(struct tdb_context *tdb, struct tdb_data key, int ltype, struct hash_info *h, - struct tdb_used_record *rec) + struct tdb_used_record *rec, + struct traverse_info *tinfo) { uint32_t i, group; tdb_off_t hashtable; @@ -148,18 +235,22 @@ tdb_off_t find_and_lock(struct tdb_context *tdb, group = use_bits(h, TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS); h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS); - /* FIXME: Guess the depth, don't over-lock! */ - h->hlock_start = (tdb_off_t)group - << (64 - (TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS)); - h->hlock_range = 1ULL << (64 - (TDB_TOPLEVEL_HASH_BITS - - TDB_HASH_GROUP_BITS)); + h->hlock_start = hlock_range(group, &h->hlock_range); if (tdb_lock_hashes(tdb, h->hlock_start, h->hlock_range, ltype, TDB_LOCK_WAIT)) return TDB_OFF_ERR; hashtable = offsetof(struct tdb_header, hashtable); + if (tinfo) { + tinfo->toplevel_group = group; + tinfo->num_levels = 1; + tinfo->levels[0].entry = 0; + tinfo->levels[0].hashtable = hashtable + + (group << TDB_HASH_GROUP_BITS) * sizeof(tdb_off_t); + tinfo->levels[0].total_buckets = 1 << TDB_HASH_GROUP_BITS; + } - while (likely(h->hash_used < 64)) { + while (h->hash_used <= 64) { /* Read in the hash group. */ h->group_start = hashtable + group * (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS); @@ -172,9 +263,23 @@ tdb_off_t find_and_lock(struct tdb_context *tdb, if (is_subhash(h->group[h->home_bucket])) { hashtable = (h->group[h->home_bucket] & TDB_OFF_MASK) + sizeof(struct tdb_used_record); + if (tinfo) { + /* When we come back, use *next* bucket */ + tinfo->levels[tinfo->num_levels-1].entry + += h->home_bucket + 1; + } group = use_bits(h, TDB_SUBLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS); h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS); + if (tinfo) { + tinfo->levels[tinfo->num_levels].hashtable + = hashtable; + tinfo->levels[tinfo->num_levels].total_buckets + = 1 << TDB_SUBLEVEL_HASH_BITS; + tinfo->levels[tinfo->num_levels].entry + = group << TDB_HASH_GROUP_BITS; + tinfo->num_levels++; + } continue; } @@ -189,15 +294,20 @@ tdb_off_t find_and_lock(struct tdb_context *tdb, if (!h->group[h->found_bucket]) break; - if (match(tdb, h, &key, h->group[h->found_bucket], rec)) + if (match(tdb, h, &key, h->group[h->found_bucket], + rec)) { + if (tinfo) { + tinfo->levels[tinfo->num_levels-1].entry + += h->found_bucket; + } return h->group[h->found_bucket] & TDB_OFF_MASK; + } } /* Didn't find it: h indicates where it would go. */ return 0; } - /* FIXME: We hit the bottom. Chain! */ - abort(); + return find_in_chain(tdb, key, hashtable, h, rec, tinfo); fail: tdb_unlock_hashes(tdb, h->hlock_start, h->hlock_range, ltype); @@ -207,8 +317,8 @@ fail: /* I wrote a simple test, expanding a hash to 2GB, for the following * cases: * 1) Expanding all the buckets at once, - * 2) Expanding the most-populated bucket, - * 3) Expanding the bucket we wanted to place the new entry ito. + * 2) Expanding the bucket we wanted to place the new entry into. + * 3) Expanding the most-populated bucket, * * I measured the worst/average/best density during this process. * 1) 3%/16%/30% @@ -283,6 +393,41 @@ int replace_in_hash(struct tdb_context *tdb, encode_offset(new_off, h)); } +/* We slot in anywhere that's empty in the chain. */ +static int COLD add_to_chain(struct tdb_context *tdb, + tdb_off_t subhash, + tdb_off_t new_off) +{ + size_t entry = tdb_find_zero_off(tdb, subhash, 1< 64) - abort(); + return add_to_chain(tdb, subhash, off); - /* FIXME: Do truncated hash bits if we can! */ h.h = hash_record(tdb, off); gnum = use_bits(&h, TDB_SUBLEVEL_HASH_BITS-TDB_HASH_GROUP_BITS); - h.group_start = subhash + sizeof(struct tdb_used_record) + h.group_start = subhash + gnum * (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS); h.home_bucket = use_bits(&h, TDB_HASH_GROUP_BITS); @@ -314,20 +457,29 @@ static int add_to_subhash(struct tdb_context *tdb, tdb_off_t subhash, static int expand_group(struct tdb_context *tdb, struct hash_info *h) { - unsigned bucket, num_vals, i; + unsigned bucket, num_vals, i, magic; + size_t subsize; tdb_off_t subhash; tdb_off_t vals[1 << TDB_HASH_GROUP_BITS]; /* Attach new empty subhash under fullest bucket. */ bucket = fullest_bucket(tdb, h->group, h->home_bucket); - subhash = alloc(tdb, 0, sizeof(tdb_off_t) << TDB_SUBLEVEL_HASH_BITS, - 0, false); + if (h->hash_used == 64) { + add_stat(tdb, alloc_chain, 1); + subsize = sizeof(struct tdb_chain); + magic = TDB_CHAIN_MAGIC; + } else { + add_stat(tdb, alloc_subhash, 1); + subsize = (sizeof(tdb_off_t) << TDB_SUBLEVEL_HASH_BITS); + magic = TDB_HTABLE_MAGIC; + } + + subhash = alloc(tdb, 0, subsize, 0, magic, false); if (subhash == TDB_OFF_ERR) return -1; - if (zero_out(tdb, subhash + sizeof(struct tdb_used_record), - sizeof(tdb_off_t) << TDB_SUBLEVEL_HASH_BITS) == -1) + if (zero_out(tdb, subhash + sizeof(struct tdb_used_record), subsize)) return -1; /* Remove any which are destined for bucket or are in wrong place. */ @@ -345,7 +497,10 @@ static int expand_group(struct tdb_context *tdb, struct hash_info *h) /* assert(num_vals); */ /* Overwrite expanded bucket with subhash pointer. */ - h->group[bucket] = subhash | ~((1ULL << (64 - TDB_OFF_UPPER_STEAL))-1); + h->group[bucket] = subhash | (1ULL << TDB_OFF_UPPER_STEAL_SUBHASH_BIT); + + /* Point to actual contents of record. */ + subhash += sizeof(struct tdb_used_record); /* Put values back. */ for (i = 0; i < num_vals; i++) { @@ -401,10 +556,6 @@ int delete_from_hash(struct tdb_context *tdb, struct hash_info *h) int add_to_hash(struct tdb_context *tdb, struct hash_info *h, tdb_off_t new_off) { - /* FIXME: chain! */ - if (h->hash_used >= 64) - abort(); - /* We hit an empty bucket during search? That's where it goes. */ if (!h->group[h->found_bucket]) { h->group[h->found_bucket] = encode_offset(new_off, h); @@ -413,6 +564,9 @@ int add_to_hash(struct tdb_context *tdb, struct hash_info *h, tdb_off_t new_off) h->group, sizeof(h->group)); } + if (h->hash_used > 64) + return add_to_chain(tdb, h->group_start, new_off); + /* We're full. Expand. */ if (expand_group(tdb, h) == -1) return -1; @@ -450,74 +604,74 @@ int add_to_hash(struct tdb_context *tdb, struct hash_info *h, tdb_off_t new_off) return add_to_hash(tdb, h, new_off); } -/* No point holding references/copies of db once we drop lock. */ -static void release_entries(struct tdb_context *tdb, - struct traverse_info *tinfo) -{ - unsigned int i; - - for (i = 0; i < tinfo->num_levels; i++) { - if (tinfo->levels[i].entries) { - tdb_access_release(tdb, tinfo->levels[i].entries); - tinfo->levels[i].entries = NULL; - } - } -} - /* Traverse support: returns offset of record, or 0 or TDB_OFF_ERR. */ static tdb_off_t iterate_hash(struct tdb_context *tdb, struct traverse_info *tinfo) { - tdb_off_t off; + tdb_off_t off, val; unsigned int i; struct traverse_level *tlevel; tlevel = &tinfo->levels[tinfo->num_levels-1]; again: - if (!tlevel->entries) { - tlevel->entries = tdb_access_read(tdb, tlevel->hashtable, - sizeof(tdb_off_t) - * tlevel->total_buckets, - true); - if (!tlevel->entries) + for (i = tdb_find_nonzero_off(tdb, tlevel->hashtable, + tlevel->entry, tlevel->total_buckets); + i != tlevel->total_buckets; + i = tdb_find_nonzero_off(tdb, tlevel->hashtable, + i+1, tlevel->total_buckets)) { + val = tdb_read_off(tdb, tlevel->hashtable+sizeof(tdb_off_t)*i); + if (unlikely(val == TDB_OFF_ERR)) return TDB_OFF_ERR; - } - /* FIXME: Use tdb_find_nonzero_off? */ - for (i = tlevel->entry; i < tlevel->total_buckets; i++) { - if (!tlevel->entries[i] || tlevel->entries[i] == tinfo->prev) + off = val & TDB_OFF_MASK; + + /* This makes the delete-all-in-traverse case work + * (and simplifies our logic a little). */ + if (off == tinfo->prev) continue; tlevel->entry = i; - off = tlevel->entries[i] & TDB_OFF_MASK; - if (!is_subhash(tlevel->entries[i])) { + if (!is_subhash(val)) { /* Found one. */ - tinfo->prev = tlevel->entries[i]; - release_entries(tdb, tinfo); + tinfo->prev = off; return off; } - /* When we come back, we want tne next one */ + /* When we come back, we want the next one */ tlevel->entry++; tinfo->num_levels++; tlevel++; tlevel->hashtable = off + sizeof(struct tdb_used_record); tlevel->entry = 0; - tlevel->entries = NULL; - tlevel->total_buckets = (1 << TDB_SUBLEVEL_HASH_BITS); + /* Next level is a chain? */ + if (unlikely(tinfo->num_levels == TDB_MAX_LEVELS + 1)) + tlevel->total_buckets = (1 << TDB_HASH_GROUP_BITS); + else + tlevel->total_buckets = (1 << TDB_SUBLEVEL_HASH_BITS); goto again; } /* Nothing there? */ - if (tinfo->num_levels == 1) { - release_entries(tdb, tinfo); + if (tinfo->num_levels == 1) return 0; + + /* Handle chained entries. */ + if (unlikely(tinfo->num_levels == TDB_MAX_LEVELS + 1)) { + tlevel->hashtable = tdb_read_off(tdb, tlevel->hashtable + + offsetof(struct tdb_chain, + next)); + if (tlevel->hashtable == TDB_OFF_ERR) + return TDB_OFF_ERR; + if (tlevel->hashtable) { + tlevel->hashtable += sizeof(struct tdb_used_record); + tlevel->entry = 0; + goto again; + } } /* Go back up and keep searching. */ - tdb_access_release(tdb, tlevel->entries); tinfo->num_levels--; tlevel--; goto again; @@ -526,7 +680,7 @@ again: /* Return 1 if we find something, 0 if not, -1 on error. */ int next_in_hash(struct tdb_context *tdb, int ltype, struct traverse_info *tinfo, - TDB_DATA *kbuf, unsigned int *dlen) + TDB_DATA *kbuf, size_t *dlen) { const unsigned group_bits = TDB_TOPLEVEL_HASH_BITS-TDB_HASH_GROUP_BITS; tdb_off_t hlock_start, hlock_range, off; @@ -549,6 +703,15 @@ int next_in_hash(struct tdb_context *tdb, int ltype, ltype); return -1; } + if (rec_magic(&rec) != TDB_USED_MAGIC) { + tdb_logerr(tdb, TDB_ERR_CORRUPT, + TDB_DEBUG_FATAL, + "next_in_hash:" + " corrupt record at %llu", + (long long)off); + return -1; + } + kbuf->dsize = rec_key_length(&rec); /* They want data as well? */ @@ -580,15 +743,57 @@ int next_in_hash(struct tdb_context *tdb, int ltype, /* Return 1 if we find something, 0 if not, -1 on error. */ int first_in_hash(struct tdb_context *tdb, int ltype, struct traverse_info *tinfo, - TDB_DATA *kbuf, unsigned int *dlen) + TDB_DATA *kbuf, size_t *dlen) { tinfo->prev = 0; tinfo->toplevel_group = 0; tinfo->num_levels = 1; tinfo->levels[0].hashtable = offsetof(struct tdb_header, hashtable); - tinfo->levels[0].entries = NULL; tinfo->levels[0].entry = 0; tinfo->levels[0].total_buckets = (1 << TDB_HASH_GROUP_BITS); return next_in_hash(tdb, ltype, tinfo, kbuf, dlen); } + +/* Even if the entry isn't in this hash bucket, you'd have to lock this + * bucket to find it. */ +static int chainlock(struct tdb_context *tdb, const TDB_DATA *key, + int ltype, enum tdb_lock_flags waitflag, + const char *func) +{ + int ret; + uint64_t h = tdb_hash(tdb, key->dptr, key->dsize); + tdb_off_t lockstart, locksize; + unsigned int group, gbits; + + gbits = TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS; + group = bits(h, 64 - gbits, gbits); + + lockstart = hlock_range(group, &locksize); + + ret = tdb_lock_hashes(tdb, lockstart, locksize, ltype, waitflag); + tdb_trace_1rec(tdb, func, *key); + return ret; +} + +/* lock/unlock one hash chain. This is meant to be used to reduce + contention - it cannot guarantee how many records will be locked */ +int tdb_chainlock(struct tdb_context *tdb, TDB_DATA key) +{ + return chainlock(tdb, &key, F_WRLCK, TDB_LOCK_WAIT, "tdb_chainlock"); +} + +int tdb_chainunlock(struct tdb_context *tdb, TDB_DATA key) +{ + uint64_t h = tdb_hash(tdb, key.dptr, key.dsize); + tdb_off_t lockstart, locksize; + unsigned int group, gbits; + + gbits = TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS; + group = bits(h, 64 - gbits, gbits); + + lockstart = hlock_range(group, &locksize); + + tdb_trace_1rec(tdb, "tdb_chainunlock", key); + return tdb_unlock_hashes(tdb, lockstart, locksize, F_WRLCK); +}